Systems, methods, and graphical user interfaces for mitigating bias in a machine learning-based decisioning model

Abstract

A system, method, and computer-program product includes obtaining a decisioning dataset comprising a plurality of favorable decisioning records and at least one unfavorable decisioning record; detecting, via a machine learning algorithm, a favorable decisioning record of the plurality of favorable decisioning records that has a vector value closest to a vector value of the unfavorable decisioning record; executing a counterfactual assessment between the favorable decisioning record and the unfavorable decisioning record; generating an explainability artifact based on one or more bias intensity metrics to explain a bias in a machine learning-based decisioning model; and in response to generating the explainability artifact, displaying the explainability artifact in a user interface.

What is claimed is: 
     
         1 . A computer-program product comprising a non-transitory machine-readable storage medium storing computer instructions that, when executed by one or more processors, perform operations comprising:
 obtaining a decisioning dataset from a machine learning-based decisioning model, wherein the decisioning dataset includes a plurality of favorable decisioning records and at least one unfavorable decisioning record;   converting, via a machine learning algorithm, the plurality of favorable decisioning records and the unfavorable decisioning record to a plurality of vector values in a multi-dimensional space;   detecting, via the machine learning algorithm, a favorable decisioning record of the plurality of favorable decisioning records that has a vector value closest to a vector value of the unfavorable decisioning record, wherein the vector value of the favorable decisioning record and the vector value of the unfavorable decisioning record are a subset of the plurality of vector values;   executing a counterfactual assessment between the favorable decisioning record and the unfavorable decisioning record, wherein the counterfactual assessment includes:
 retrieving attributes of the favorable decisioning record and the unfavorable decisioning record, and 
 computing one or more bias intensity metrics based on disparities between the attributes of the favorable decisioning record and the unfavorable decisioning record; 
   generating an explainability artifact that uses the one or more bias intensity metrics to explain a bias in the machine learning-based decisioning model, wherein explaining the bias in the machine learning-based decisioning model includes visually emphasizing that the machine learning-based decisioning model is generating unfair decisions; and   in response to generating the explainability artifact, displaying the explainability artifact in a user interface that is accessible by a user, wherein the explainability artifact enables the user to execute a bias mitigation process that reconfigures one or more components of the machine learning-based decisioning model to mitigate the bias in the machine learning-based decisioning model.   
     
     
         2 . The computer-program product according to  claim 1 , wherein:
 a first bias intensity metric of the one or more bias intensity metrics represents a difference between a respective attribute of the favorable decisioning record and the unfavorable decisioning record,   the explainability artifact explains the bias in the machine learning-based decisioning model by indicating factors causing the bias in the machine learning-based decisioning model, and   indicating the factors causing the bias in the machine learning-based decisioning model includes:
 determining that a value of the first bias intensity metric is less than a threshold amount, and 
 based on determining that the value of the first bias intensity metric is less than the threshold amount, displaying, in the explainability artifact, the first bias intensity metric with visual emphasis indicating the respective attribute associated with the first bias intensity metric is a factor contributing to the bias in the machine learning-based decisioning model. 
   
     
     
         3 . The computer-program product according to  claim 2 , wherein:
 a second bias intensity metric of the one or more bias intensity metrics represents a difference between a second respective attribute of the favorable decisioning record and the unfavorable decisioning record, and   indicating the factors causing the bias in the machine learning-based decisioning model further includes:
 determining that a value of the second bias intensity metric is more than the threshold amount, and 
 based on determining that the value of the second bias intensity metric is more than the threshold amount, displaying, in the explainability artifact, the second bias intensity metric with visual emphasis indicating the second respective attribute associated with the second bias intensity metric is not the factor contributing to the bias in the machine learning-based decisioning model. 
   
     
     
         4 . The computer-program product according to  claim 1 , wherein:
 the one or more bias intensity metrics includes a first bias intensity metric that represents a difference between a respective attribute of the favorable decisioning record and the unfavorable decisioning record, and   using the one or more bias intensity metrics to explain the bias in the machine learning-based decisioning model includes:
 displaying, in the explainability artifact, a value of the first bias intensity metric, and 
 indicating, via visual emphasis displayed in association with the value of the first bias intensity metric, an unfairness of the machine learning-based decisioning model when the value of the first bias intensity metric indicates that the difference between the respective attribute of the favorable decisioning record and the unfavorable decisioning record is less than a threshold amount. 
   
     
     
         5 . The computer-program product according to  claim 4 , wherein:
 using the one or more bias intensity metrics to explain the bias in the machine learning-based decisioning model further includes:
 indicating, via the visual emphasis displayed in association with the value of the first bias intensity metric, a fairness of the machine learning-based decisioning model when the value of the first bias intensity metric indicates that the difference between the respective attribute of the favorable decisioning record and the unfavorable decisioning record is more than the threshold amount. 
   
     
     
         6 . The computer-program product according to  claim 1 , wherein:
 the machine learning algorithm is configurable to detect a second plurality of favorable decisioning records that have vector values closest to the vector value of the unfavorable decisioning record, and   detecting, via the machine learning algorithm, the second plurality of favorable decisioning records that have the vector values closest to the vector value of the unfavorable decisioning record at least includes:
 detecting the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record, and 
 detecting a second favorable decisioning record of the plurality of favorable decisioning records that has a vector value that is second closest to the vector value of the unfavorable decisioning record. 
   
     
     
         7 . The computer-program product according to  claim 1 , wherein:
 the machine learning algorithm computes a plurality of vector distances between the unfavorable decisioning record and the plurality of favorable decisioning records, and   detecting, via the machine learning algorithm, the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record includes:
 identifying a smallest vector distance among the plurality of vector distances computed by the machine learning algorithm, wherein a vector distance between the unfavorable decisioning record and a first favorable decisioning record of the plurality of favorable decisioning records is identified as having the smallest vector distance, and 
 selecting the first favorable decisioning record as the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record. 
   
     
     
         8 . The computer-program product according to  claim 7 , wherein a respective vector distance of the plurality of vector distances computed by the machine learning algorithm denotes a Euclidean distance between the vector value of the unfavorable decisioning record and a vector value of a respective favorable decisioning record. 
     
     
         9 . The computer-program product according to  claim 1 , wherein:
 the machine learning algorithm detects the favorable decisioning record and further detects a second favorable decisioning record of the plurality of favorable decisioning records that has a vector value that is second closest to the vector value of the unfavorable decisioning record,   the computer instructions, when executed by the one or more processors, perform operations further comprising:
 retrieving the attributes of the second favorable decisioning record; and 
 computing one or more second bias intensity metrics based on disparities between the attributes of the unfavorable decisioning record and the attributes of the second favorable decisioning record, and 
   the explainability artifact uses the one or more bias intensity metrics and further uses the one or more second bias intensity metrics to explain the bias in the machine learning-based decisioning model.   
     
     
         10 . The computer-program product according to  claim 1 , wherein:
 the decisioning dataset includes a plurality unfavorable decisioning records, including the unfavorable decisioning record and one or more other unfavorable decisioning records, and   the computer instructions, when executed by the one or more processors, perform operations further comprising:
 executing a plurality of counterfactual assessments, wherein executing the plurality of counterfactual assessments includes:
 executing the counterfactual assessment between the favorable decisioning record and the unfavorable decisioning record, and 
 executing one or more additional counterfactual assessments between the one or more other unfavorable decisioning records and one or more of the plurality of favorable decisioning records. 
 
   
     
     
         11 . The computer-program product according to  claim 1 , wherein a respective bias intensity metric of the one or more bias intensity metrics represents a difference between a respective attribute of the unfavorable decisioning record and the favorable decisioning record. 
     
     
         12 . The computer-program product according to  claim 1 , wherein:
 the decisioning dataset includes one or more attributes, including an account score attribute, and   computing the one or more bias intensity metrics includes computing an account score bias intensity metric that represents a difference between the account score attribute of the unfavorable decisioning record and the account score attribute of the favorable decisioning record.   
     
     
         13 . The computer-program product according to  claim 1 , wherein:
 the decisioning dataset includes one or more attributes, including a resource attribute, and   computing the one or more bias intensity metrics includes computing a resource bias intensity metric that represents a difference between the resource attribute of the unfavorable decisioning record and the resource attribute of the favorable decisioning record.   
     
     
         14 . The computer-program product according to  claim 1 , wherein:
 the machine learning-based decisioning model comprises one or more sub-models, including at least one heuristic-based model comprising a plurality of decisioning heuristics, and   reconfiguring the one or more components of the machine learning-based decisioning model to mitigate the bias includes:
 reconfiguring thresholds of the plurality of decisioning heuristics to mitigate the heuristic-based model from generating unfavorable decisions for decisioning records that have similar attributes to the unfavorable decisioning record. 
   
     
     
         15 . The computer-program product according to  claim 1 , wherein:
 the machine learning-based decisioning model comprises one or more sub-models, including at least one machine learning model, and   reconfiguring the one or more components of the machine learning-based decisioning model to mitigate the bias includes:
 updating a weight and a bias of the machine learning model to mitigate the machine learning model from generating unfavorable decisions for decisioning records that have similar attributes to the unfavorable decisioning record. 
   
     
     
         16 . The computer-program product according to  claim 1 , wherein:
 the machine learning-based decisioning model comprises one or more sub-models, including at least one machine learning model, and   reconfiguring the one or more components of the machine learning-based decisioning model to mitigate the bias includes:
 reconfiguring an algorithmic structure of the machine learning model to include model constraints that constrain the machine learning model from generating unfavorable decisions for decisioning records that have similar attributes to the unfavorable decisioning record. 
   
     
     
         17 . The computer-program product according to  claim 1 , wherein:
 the machine learning-based decisioning model comprises one or more sub-models, including at least one machine learning model, and   reconfiguring the one or more components of the machine learning-based decisioning model to mitigate the bias includes:
 transforming a prediction by the machine learning model from an unfavorable prediction to a favorable prediction when a corresponding decisioning record has factors causing the bias. 
   
     
     
         18 . The computer-program product according to  claim 1 , wherein the machine learning algorithm uses a k-nearest neighbor algorithm to detect the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record. 
     
     
         19 . The computer-program product according to  claim 1 , wherein the explainability artifact comprises a plurality of data visualization components, including:
 a first table component that includes a row corresponding to the unfavorable decisioning record and columns corresponding to the attributes of the unfavorable decisioning record,   a second table component that includes a row corresponding to the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record and columns corresponding to the one or more bias intensity metrics, and   a graph component that graphs the unfavorable decisioning record and one or more other unfavorable decisioning records in the decisioning dataset to illustrate a distribution of a respective attribute shared by the unfavorable decisioning record and the one or more other unfavorable decisioning records.   
     
     
         20 . The computer-program product according to  claim 19 , wherein:
 the machine learning algorithm detects the favorable decisioning record and further detects a second favorable decisioning record of the plurality of favorable decisioning records that has a vector value that is second closest to the vector value of the unfavorable decisioning record, and   the second table component includes the row corresponding to the favorable decisioning record and further includes a second row corresponding to the second favorable decisioning record that has the vector value that is second closest to the vector value of the unfavorable decisioning record.   
     
     
         21 . The computer-program product according to  claim 1 , wherein explaining the bias in the machine learning-based decisioning model includes surfacing at least one of the attributes contributing to the bias in the machine learning-based decisioning model. 
     
     
         22 . The computer-program product according to  claim 1 , wherein explaining the bias in the machine learning-based decisioning model includes surfacing the disparities between the attributes of the unfavorable decisioning record and the favorable decisioning record that has the vector value closest to the vector value of the unfavorable decisioning record. 
     
     
         23 . The computer-program product according to  claim 1 , wherein the explainability artifact displays the attributes of the unfavorable decisioning record concurrently with entries that indicate:
 the disparities between the attributes of the unfavorable decisioning record and the favorable decisioning record that has the vector value closest to the unfavorable decisioning record, and   second disparities between the attributes of the unfavorable decisioning record and one or more other favorable decisioning records of the plurality of favorable decisioning records that have vector values second closest to the unfavorable decisioning record.   
     
     
         24 . A computer-implemented method comprising:
 obtaining a decisioning dataset from a machine learning-based decisioning model, wherein the decisioning dataset includes a plurality of favorable decisioning records and at least one unfavorable decisioning record;   converting, via a machine learning algorithm, the plurality of favorable decisioning records and the unfavorable decisioning record to a plurality of vector values in a multi-dimensional space;   detecting, via the machine learning algorithm, a favorable decisioning record of the plurality of favorable decisioning records that has a vector value closest to a vector value of the unfavorable decisioning record, wherein the vector value of the favorable decisioning record and the vector value of the unfavorable decisioning record are a subset of the plurality of vector values;   executing a counterfactual assessment between the favorable decisioning record and the unfavorable decisioning record, wherein the counterfactual assessment includes:
 retrieving attributes of the favorable decisioning record and the unfavorable decisioning record, and 
 computing one or more bias intensity metrics based on disparities between the attributes of the favorable decisioning record and the unfavorable decisioning record; 
   generating an explainability artifact that uses the one or more bias intensity metrics to explain a bias in the machine learning-based decisioning model, wherein explaining the bias in the machine learning-based decisioning model includes visually emphasizing that the machine learning-based decisioning model is generating unfair decisions; and   in response to generating the explainability artifact, displaying the explainability artifact in a user interface that is accessible by a user, wherein the explainability artifact enables the user to execute a bias mitigation process that reconfigures one or more components of the machine learning-based decisioning model to mitigate the bias in the machine learning-based decisioning model.   
     
     
         25 . The computer-implemented method according to  claim 24 , wherein:
 a first bias intensity metric of the one or more bias intensity metrics represents a difference between a respective attribute of the favorable decisioning record and the unfavorable decisioning record,   the explainability artifact explains the bias in the machine learning-based decisioning model by indicating factors causing the bias in the machine learning-based decisioning model, and   indicating the factors causing the bias in the machine learning-based decisioning model includes:
 determining that a value of the first bias intensity metric is less than a threshold amount, and 
 based on determining that the value of the first bias intensity metric is less than the threshold amount, displaying, in the explainability artifact, the first bias intensity metric with visual emphasis indicating the respective attribute associated with the first bias intensity metric is a factor contributing to the bias in the machine learning-based decisioning model. 
   
     
     
         26 . The computer-implemented method according to  claim 25 , wherein:
 a second bias intensity metric of the one or more bias intensity metrics represents a difference between a second respective attribute of the favorable decisioning record and the unfavorable decisioning record, and   indicating the factors causing the bias in the machine learning-based decisioning model further includes:
 determining that a value of the second bias intensity metric is more than the threshold amount, and 
   
       based on determining that the value of the second bias intensity metric is more than the threshold amount, displaying, in the explainability artifact, the second bias intensity metric with visual emphasis indicating the second respective attribute associated with the second bias intensity metric is not the factor contributing to the bias in the machine learning-based decisioning model. 
     
     
         27 . A computer-implemented system comprising:
 one or more processors;   a memory; and   a computer-readable medium operably coupled to the one or more processors, the computer-readable medium having computer-readable instructions stored thereon that, when executed by the one or more processors, cause a computing device to perform operations comprising:
 obtaining a decisioning dataset from a machine learning-based decisioning model, wherein the decisioning dataset includes a plurality of favorable decisioning records and at least one unfavorable decisioning record; 
 converting, via a machine learning algorithm, the plurality of favorable decisioning records and the unfavorable decisioning record to a plurality of vector values in a multi-dimensional space; 
 detecting, via the machine learning algorithm, a favorable decisioning record of the plurality of favorable decisioning records that has a vector value closest to a vector value of the unfavorable decisioning record, wherein the vector value of the favorable decisioning record and the vector value of the unfavorable decisioning record are a subset of the plurality of vector values; 
 executing a counterfactual assessment between the favorable decisioning record and the unfavorable decisioning record, wherein the counterfactual assessment includes:
 retrieving attributes of the favorable decisioning record and the unfavorable decisioning record, and 
 computing one or more bias intensity metrics based on disparities between the attributes of the favorable decisioning record and the unfavorable decisioning record; 
 
 generating an explainability artifact that uses the one or more bias intensity metrics to explain a bias in the machine learning-based decisioning model, wherein explaining the bias in the machine learning-based decisioning model includes visually emphasizing that the machine learning-based decisioning model is generating unfair decisions; and 
 in response to generating the explainability artifact, displaying the explainability artifact in a user interface that is accessible by a user, wherein the explainability artifact enables the user to execute a bias mitigation process that reconfigures one or more components of the machine learning-based decisioning model to mitigate the bias in the machine learning-based decisioning model. 
   
     
     
         28 . The computer-implemented system according to  claim 27 , wherein:
 the one or more bias intensity metrics includes a first bias intensity metric that represents a difference between a respective attribute of the favorable decisioning record and the unfavorable decisioning record, and   using the one or more bias intensity metrics to explain the bias in the machine learning-based decisioning model includes:
 displaying, in the explainability artifact, a value of the first bias intensity metric, and 
 indicating, via visual emphasis displayed in association with the value of the first bias intensity metric, an unfairness of the machine learning-based decisioning model when the value of the first bias intensity metric indicates that the difference between the respective attribute of the favorable decisioning record and the unfavorable decisioning record is less than a threshold amount. 
   
     
     
         29 . The computer-implemented system according to  claim 28 , wherein:
 using the one or more bias intensity metrics to explain the bias in the machine learning-based decisioning model further includes:
 indicating, via the visual emphasis displayed in association with the value of the first bias intensity metric, a fairness of the machine learning-based decisioning model when the value of the first bias intensity metric indicates that the difference between the respective attribute of the favorable decisioning record and the unfavorable decisioning record is more than the threshold amount. 
   
     
     
         30 . The computer-implemented system according to  claim 27 , wherein a respective bias intensity metric of the one or more bias intensity metrics represents a difference between a respective attribute of the unfavorable decisioning record and the favorable decisioning record.